Can Steam Reviews Estimate Sales? Common Indie Research Mistakes
Reviews can support sales scenarios, but reviewing behavior differs between games. Establish review filters and product stage before applying any multiplier.
Key findings
- Reviews are not purchase orders.
- Total, purchase-only and language-filtered reviews are not interchangeable.
- A model’s predictions are not ground-truth training labels.
Why reviews are used to infer sales
Many Steam games do not continually disclose sales, while reviews are visible signals. Buyers who choose to review are only part of the buying audience. Credible historical samples can help describe that relationship and support sales scenarios. That is the practical attraction of review-based estimation.
The relationship varies with the game, audience and time. Social discussion, controversy and particularly committed players can change how often reviews appear. A ratio that works for one title should not be assumed to represent the entire storefront.
Establish which reviews the calculation uses
The Steam review API supports purchase-type, language and other filters. English-only reviews differ from reviews in every language, and all-source reviews differ from Steam-purchase reviews. Inconsistent definitions do not become comparable just because a calculation displays many decimal places.
Record the filters alongside the count before selecting an industry multiplier. Language coverage is especially important when comparing games with different regional audiences: it changes which reviews enter the observation before any sales assumptions are applied.
A sales scenario is only the first revenue step
Even with a range for estimated copies, multiplying by today’s list price is insufficient. Historical discounts, regional transaction mix and refunds affect average spending. Uncertainty in the sales estimate continues into the revenue result rather than disappearing at the next multiplication.
A useful output explains how assumptions affect the answer. Check whether relative positions remain stable between conservative and higher scenarios. If a small assumption change reverses the ordering, a highly confident ranking is not warranted by the evidence.
How a model can improve over time
Prefer developer disclosures or authorized sales records as training material, aligned to a common date and channel. Treating an earlier model estimate as a true sales label merely reinforces the model’s existing assumptions. It does not add independent evidence.
Preserve predictions and facts together but present them separately. Reviews are observations, official disclosures are facts within a particular scope, and a sales model is an inference. Keeping those categories clear allows later disclosures to expose errors instead of perpetuating a difficult-to-check number.
Sources
Explore these games
Explore more history and deeper comparisons.
View plans